How Deterministic Tools Took AI Biology From 16.9% to 92.8%
Anthropic’s AI for Science Event Revealed the Single Change That Closed a Massive Accuracy Gap — and Claude Science Is Now Pointing That Capability at Neglected Diseases No Pharmaceutical Company Has a Financial Reason to Study.
At Anthropic’s AI for Science event on June 30th, the company revealed a finding with a bigger real-world impact than most product launches this series has covered: pairing Claude with deterministic computational tools instead of letting it answer biology questions from language alone pushed accuracy from 16.9% to 92.8% on a benchmark task. That gap is the entire reason Claude Science exists, and the entire reason Anthropic can now credibly point AI at drug discovery for diseases the pharmaceutical industry has no financial incentive to touch. Day 33 explains what a deterministic tool actually is, and why that one design choice mattered more than any model upgrade.
16.9% is barely better than a coin flip on a well-defined biology task. 92.8% is genuinely useful. The distance between those two numbers is not a bigger model, more training data, or a longer context window. It is one architectural decision: stop asking the language model to compute an answer from memory, and instead let it call a tool that computes the answer correctly every time. That single change is quietly one of the most important developments in applied AI this year.
Neal Lloyd · Inside The Machine, Day 33Ground Truth, Episode 23, closed out the fullest account yet of the Fable 5 ban and covered two new business stories breaking the same week. This series has a quieter but arguably more consequential story from earlier in July to finish properly: at Anthropic’s AI for Science event on June 30th, alongside the Claude Science workbench launch, the company disclosed that a specific architectural change — giving Claude access to deterministic computational tools rather than asking it to answer biology questions from language alone — moved accuracy on a benchmark task from 16.9% to 92.8%. John Jumper, the AlphaFold lead Anthropic hired from Google DeepMind, brought exactly the kind of protein-structure expertise that informed how that toolkit was built. This is Day 33 of Inside The Machine. Today we explain what a deterministic tool actually is, why the gap it closes is so large, and what Anthropic is now pointing the resulting capability at.
The Difference Between Guessing an Answer and Computing One
A language model answering a biology question purely from its training data is doing something closer to sophisticated recall than calculation: it has seen enough text about protein structures, molecular interactions, and biological mechanisms to produce plausible-sounding answers, but it is not actually running the underlying physics or chemistry. That works reasonably well for questions with well-documented answers already in the training data. It works badly for questions that require genuine computation — predicting how a specific novel molecule folds, for instance, is not something a model can reliably reconstruct from memory alone, because the correct answer for a truly novel input was never in the training data to begin with.
A deterministic tool is different in kind, not degree. Rather than asking the model to produce the answer directly, the system routes the question to a purpose-built computational tool — a structural biology solver, a chemistry simulator, a database lookup against verified experimental data — that returns the same correct answer every time for the same input, the way a calculator returns 4 for 2+2 regardless of who asks. The model’s job shifts from generating the answer to correctly deciding which tool to call, feeding it the right inputs, and interpreting the output.
The 60-plus tools integrated into the Claude Science workbench include exactly this kind of infrastructure, with the AlphaFold Database — the protein structure repository behind the 2024 Nobel Prize-winning work John Jumper led at DeepMind before joining Anthropic — as one of the most heavily used sources. Jumper’s specific expertise in protein structure informatics is precisely the domain knowledge that shaped how the toolkit routes and validates biological queries.
16.9%: Claude’s accuracy on the benchmark biology task using language-only reasoning, barely above random guessing on a task with several plausible answer categories. 92.8%: accuracy on the same task once deterministic tools were added to the pipeline. 60-plus: computational tools integrated into the Claude Science workbench, spanning structural biology, chemistry, and verified experimental databases.
Language Models Are Built to Sound Right, Not to Be Right
A large language model is trained to predict plausible next tokens based on patterns in its training data — it is optimised, fundamentally, for producing text that looks like a correct answer to the kind of question being asked. For many domains, looking correct and being correct converge closely enough that the distinction rarely matters in practice. Biology, and structural biology in particular, is an unusually bad match for that optimisation target: a plausible-sounding protein interaction that is subtly wrong is often indistinguishable from a correct one to a non-expert reader, and the training data itself contains a mix of well-established science, contested hypotheses, and outright errors that a model has no principled way to weight correctly on its own.
Deterministic tools sidestep the problem entirely rather than trying to solve it through better training. A structural biology solver does not produce a plausible-sounding fold prediction — it runs an actual computation grounded in physical and chemical constraints, and returns the same result every time for the same input. The 92.8% figure is not the model getting smarter at biology; it is the model getting better at recognising when a question needs a tool rather than a guess, and correctly routing to that tool instead of answering from its own uncertain internal knowledge.
That distinction matters well beyond this one benchmark. It is the same underlying principle behind why a calculator embedded in a chat interface is trusted for arithmetic in a way the base model alone is not, and why this series has repeatedly flagged hallucination as a risk specifically in domains where a model is asked to compute rather than recall. Claude Science is, in a real sense, a bet that the biggest near-term gains in applied AI for science come from routing more of biology’s hard questions to tools like this, rather than from larger models memorising more of the literature.
A 92.8% accuracy score sounds like a model got smarter. What actually happened is closer to a model learning when to stop guessing. That is a less flattering story for the marketing, and a far more useful one for anyone deciding whether to trust the answer in front of them.Neal Lloyd · Inside The Machine, Day 33
Neglected Diseases Are Neglected Because of Economics, Not Difficulty
Alongside the Claude Science workbench launch, Anthropic announced an internal drug discovery program specifically targeting neglected diseases — conditions that disproportionately affect low-income populations and that major pharmaceutical companies have limited financial incentive to research, because the patient populations able to pay commercial drug prices are small relative to the cost of running a full discovery pipeline. This is the textbook market failure in pharmaceutical research: the diseases causing the most preventable suffering are often the ones with the weakest commercial case for treatment development.
Deterministic-tool-augmented AI changes the economics of at least the earliest, most expensive stage of that pipeline: computational screening and structural prediction. If a research team can screen candidate compounds against a target protein computationally, with genuine 92.8%-tier accuracy rather than language-model guesswork, at a fraction of the cost of wet-lab experimentation, the threshold for a neglected-disease program to be worth funding — philanthropically or through smaller research grants — drops substantially. Anthropic is pairing the internal program with a Claude Science AI for Science grants initiative offering $30,000 in credits to 50 research projects, with applications closing July 15th, extending access specifically to academic and independent researchers who would not otherwise have workbench-tier resources.
None of this guarantees a treatment gets developed for any specific neglected disease — computational screening is one early stage in a long, expensive pipeline that still requires wet-lab validation, clinical trials, and manufacturing at every subsequent step. What it does is lower the cost of the stage most likely to be skipped entirely for lack of funding. Whether that translates into treatments reaching patients who currently have none is a multi-year question this series will be positioned to revisit.
Neglected diseases are not scientifically harder than the ones that get funded. They are commercially unrewarding. A tool that makes the cheapest, most skippable stage of drug discovery meaningfully cheaper does not fix that incentive problem — but it does lower the bar for someone to fund the work anyway, which is a real, if modest, kind of progress.Neal Lloyd · Inside The Machine, Day 33
Inside The Machine, Day 33 · July 11 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs Closed
- Day 15The New Cold War
- Day 16Why AI Lies With Confidence
- Day 17AI Is Eating the Power Grid
- Day 18The Age of AI Agents
- Day 19AI Safety Was Never Just Theory
- Day 20The Surveillance Question
- Day 21AI and the Future of Education
- Day 22AI and Your Health
- Day 23What Is AGI and Are We Close?
- Day 24What Is Work For?
- Day 25AI and Democracy
- Day 26AI and the Future of Money
- Day 27Can the Planet Afford AI?
- Day 28Why AI Forgets Everything
- Day 29Can Anyone Actually Govern AI Now?
- Day 30Inside the Jailbreak Severity Framework
- Day 31Why No One Can Guarantee Your AI Agent Will Do What It Was Told
- Day 32Squidbleed: A 29-Year-Old Bug and the Same Capability That Got Fable 5 Recalled
- Day 33How Deterministic Tools Took AI Biology From 16.9% to 92.8%You are here
